Koliousis, Alexandros K.
An elementary proposition on the dynamic routing
problem in wireless networks of sensors.
PhD thesis, University of Glasgow.
Full text available as:
The routing problem (finding an optimal route from one point in a computer network to another) is surrounded by impossibility results. These results are usually expressed as lower and upper bounds on the set of nodes (or the set of links) of a network and represent the complexity of a solution to the routing problem (a routing function).
The routing problem dealt with here, in particular, is a dynamic one (it accounts for network dynamics) and concerns wireless networks of sensors. Sensors form wireless links of limited capacity and time-variable quality to route messages amongst themselves. It is desired that sensors self-organize ad hoc in order to successfully carry out a routing task, e.g. provide daily soil erosion reports for a monitored watershed, or provide immediate indications of an imminent volcanic eruption, in spite of network dynamics.
Link dynamics are the first barrier to finding an optimal route between a node x and a node y in a sensor network. The uncertainty of the outcome (the best next hop) of a routing function lies partially with the quality fluctuations of wireless links. Take, for example, a static network. It is known that, given the set of nodes and their link weights (or costs), a node can compute optimal routes by running, say, Dijkstra's algorithm. Link dynamics however suggest that costs are not static. Hence, sensors need a metric (a measurable quantity of uncertainty) to monitor for fluctuations, either improvements or degradations of quality or load; when a fluctuation is sufficiently large (say, by Delta), sensors ought to update their costs and seek another route. Therein lies the other fundamental barrier to find an optimal route - complexity.
A crude argument would suggest that sensors (and their links) have an upper bound on the number of messages they can transmit, receive and store due to resource constraints. Such messages can be application traffic, in which case it is desirable, or control traffic, in which case it should be kept minimal. The first type of traffic is demand, and a user should provision for it accordingly. The second type of traffic is overhead, and it is necessary if a routing system (or scheme) is to ensure its fidelity to the application requirements (policy). It is possible for a routing scheme to approximate optimal routes (by Delta) by reducing its message and/or memory complexity.
The common denominator of the routing problem and the desire to minimize overhead while approximating optimal routes is Delta, the deviation (or stretch) of a computed route from an optimal one, as computed by a node that has instantaneous knowledge of the set of all nodes and their interaction costs (an oracle). This dissertation deals with both problems in unison. To do so, it needs to translate the policy space (the user objectives) into a metric space (routing objectives). It does so by means of a cost function that normalizes metrics into a number of hops. Then it proceeds to devise, design, and implement a scheme that computes minimum-hop-count routes with manageable complexity.
The theory presented is founded on (well-ordered) sets with respect to an elementary proposition, that a route from a source x to a destination y can be computed either by y sending an advertisement to the set of all nodes, or by x sending a query to the set of all nodes; henceforth the proactive method (of y) and the reactive method (of x), respectively.
The debate between proactive and reactive routing protocols appears in many instances of the routing problem (e.g. routing in mobile networks, routing in delay-tolerant networks, compact routing), and it is focussed on whether nodes should know a priori all routes and then select the best one (with the proactive method), or each node could simply search for a (hopefully best) route on demand (with the reactive method).
The proactive method is stateful, as it requires the entire metric space - the set of nodes and their interaction costs - in memory (in a routing table). The routes computed by the proactive method are optimal and the lower and upper bounds of proactive schemes match those of an oracle. Any attempt to reduce the proactive overhead, e.g. by introducing hierarchies, will result in sub-optimal routes (of known stretch). The reactive method is stateless, as it requires no information whatsoever to compute a route. Reactive schemes - at least as they are presently understood - compute sub-optimal routes (and thus far, of unknown stretch).
This dissertation attempts to answer the following question: "what is the least amount of state required to compute an optimal route from a source to a destination?" A hybrid routing scheme is used to investigate this question, one that uses the proactive method to compute routes to near destinations and the reactive method for distant destinations.
It is shown that there are cases where hybrid schemes can converge to optimal routes, despite possessing incomplete routing state, and that the necessary and sufficient condition to compute optimal routes with local state alone is related neither to the size nor the density of a network; it is rather the circumference (the size of the largest cycle) of a network that matters. Counterexamples, where local state is insufficient, are discussed to derive the worst-case stretch.
The theory is augmented with simulation results and a small experimental testbed to motivate the discussion on how policy space (user requirements) can translate into metric spaces and how different metrics affect performance. On the debate between proactive and reactive protocols, it is shown that the two classes are equivalent. The dissertation concludes with a discussion on the applicability of its results and poses some open problems.
Actions (login required)